Adolf Knoll, National Library of the Czech Republic, Prague
In most cases, document delivery services as well as various digitization activities are based on black-and-white images. These digital images can provide access to most important textual and partly also graphical information contained in newspapers, journals, and other types of printed documents and modern records. This is the information, which is not available in electronic form, because of the fact that it was printed on paper, but its digitization can contribute to its wider dissemination and also more manageable archiving.
Let us imagine a page from an old journal: its original size was ca. 83 x 91,5 cm, it was scanned from microfilm at 150 dpi. Thus, we need for it 4901 x 5410 = 26514410 pixels that means for true colour image - in which each pixel consists of 24 bits = 3 bytes - ca. 79,5 MB. If we take in consideration that we would like to provide Internet access to such an image or to image of pages of this journal, it would be very difficult to transfer approximately 80 MB of data for each page. For much Internet access in slow networks it would be even impossible.
The old journals consisted mostly of black texts on white pages with scarce drawings and later also with photographs. For most of them the black-and-white image (bi-level or bi-tonal image) would be good enough to offer appropriate access to the information contained.
The bi-level image means that for each pixel (picture element) from the scanned image we do not need three bytes as in the true colour image (over 16.7 million colours), but only one bit that is 1/24 part of the pixel size from the true colour image. The one-bit pixel can express only black or white colours, nothing more, but if the number of pixels per unit of length (called resolution) is sufficiently higher, such a solution can bring satisfactory results. In case of our above-mentioned page, the size of the corresponding bi-level image will be ca. 3.314 MB. However, even such a file is pretty large for reasonable transfer; therefore, techniques must be found to compress it down.
Bi-level images are a very classical domain of interest of fax industry, because facsimile machines need efficient standards for speeding up transmission of documents. In 1980, an international standard called Group 3 was recommended by a study group of the International Telegraph and Telephone Consultative Committee (CCITT) for this purpose. This standard implemented a simple compression scheme that enabled a substantial lossless reduction of transmitted bi-level files. If applied to our file, the result will be 477 KB of data that is more manageable than the uncompressed form. Most facsimile machines of today use still this compression scheme for document transfer.
However, further work on this CCITT Group 3 standard went on, and in 1984 an improved compression scheme called Group 4 was recommended as an international standard. If not very much used in daily faxing, the standard found its way in digital imaging as well as the previous CCITT G3 standard. This happened thanks to the fact that both of them were incorporated in the TIFF 4.0 format.
In the case of our page from the old journal the G4 scheme is able to compress the file down onto 192 KB without losses of information face to the uncompressed file. This result is not bad compared to 3314 KB of the uncompressed file; therefore, the G4 scheme implemented in TIFF is also frequently used as the output and access format for most scanned documents. It was also recommended for the textual documents digitized within the UNESCO Memory of the World programme.
Nevertheless, the search of better and better compression schemes for the bi-level image went on. In 1993, the Joint Bi-level Image Experts Group (JBIG) produced a new bi-level image coding standard (ISO/IEC 11544) known as JBIG and called nowadays also JBIG1. However, this compression scheme, which outperformed the G4 standard, was not largely applied, because there was no commonly used data format enabling the practical work with this scheme.
Nowadays, it is expected that an improved JBIG2 compression scheme will be adopted as an international ISO standard during the year 2000, because it has already got the Final Draft ISO International Standard status. The JBIG2 scheme will enable further lossless and even lossy compression of the bi-level image.
There are several companies, which are experimenting now with implementation of JBIG2 into usable data formats, for example, the Xerox Corp. and Image Power Inc. The latter company has enabled it in its recent product called Power Compressor1 as the file type JB2.
Other companies are following similar principles to elaborate efficient compression schemes for the bi-level image. The AT&T laboratories implemented their variation proposal for JBIG2 into their new DJVU format as a JB2 sub-component for encoding that part of the image that is considered to be foreground. The German LuraTech Comp. designed another bi-level encoding scheme that was implemented in their new LDF format for the same reason as it was done with JB2 by the AT&T laboratories. This scheme has no specific name; therefore, it will be referred to as LDF bi-tonal encoder. Finally, there is also a Lightning Strike solution developed by Infinop, Inc.
All these schemes - and possibly also other ones - have in common further tightening of compression. Some of them are well elaborated and it is possible to test them, while the authoring or viewing tools for other ones are not available. Usually, the new schemes offer both lossless and lossy compression. Thus, we achieved the best results for our journal page through application of the DJVU JB2 encoder: 154 KB for the lossless option and 109 KB for the so-called aggressive lossy option whose quality was, however, fully acceptable. Compared with 192 KB of CCITT G4 this is a very good result as we can see below on the Fig. no. 1.
Fig. no. 1 – Comparison of file sizes after compression by various schemes (old journal)
Of course, even more interesting would be to compare the 109 KB achieved by JB2 in DJVU with 79,5 MB for the true-colour image as shown in Fig. no. 2. Thus, it is quite evident that the bi-level image has its good future and that the search of better compression schemes - the lossy ones included - can contribute to easier, cheaper, and faster access to scanned information.
The advantages of such solutions have been shown on a rather large file produced through digitization of an old 19th century journal. Keeping in view the processing techniques to enable such a compression ratio, it becomes evident that compression of more recent texts and better-quality documents should give far better results.
As we are engaged in a digitization programme related to old newspapers and journals, it was rather tentative for us to get better orientation in this problem and to test available schemes and tools.
The DJVU and LDF formats have, however, another philosophy. They were designed to enable the so-called mixed raster content encoding through combining of different schemes for different image layers. Both of them split the image between foreground and background layers, while the foreground layers are encoded in bi-level schemes: in DJVU it is a JB2 scheme and in LDF there is an option between CCITT G4 and the LDF own bi-tonal encoder. The foreground level is normally combined with background layers encoded by a modern wavelet compression technique called IW44 for DJVU and LWF for LDF formats. One of these background layers is in fact applicable as a background image for the bi-level foreground image, while the other layers (one in LDF and four in IW44) are used for images.
The foreground layer concerns the information that is considered to be the text as the encoding scheme is based on repeatability of characters or pixel groups on the page image and the sophisticated encoding of the found redundancy.
The problem was, whether the above mentioned authoring tools would enable solely encoding of the bi-level image without application of wavelets for background. The DjVu™ Shop enables it directly, because it offers the option to process any image as a black-and-white document. In this way, the DJVU format contains only the JB2 component, while the IW44 components are switched off. The JB2 bi-level encoding can be set as lossless, lossy, or aggressive. The file extension even in this case remains to be DJVU (e.g., myfile.djvu). It is accessible in web browsers (both Netscape and Internet Explorer) with the help of a user-friendly plug-in.
In the case of the LDF bi-tonal encoder, however, the situation is more complicated as LDF does not enable the direct encoding of bi-level source files. The solution is to increase the number of colours up to 16 million and to set the option in the dialogue menu that everything should be encoded and that this everything should be considered to be text. If so, the LDF Adobe Photoshop Plug-In applies really only its bi-level encoder, but the manipulation with images with increased number of colours requires much computing.
If the user-friendliness of encoding software combined with its speed and computation efficiency is evaluated, the winner is the DjVu™ Shop. It has never happened - even in case of very large files - larger than that mentioned beforehand - that the encoder or the computer have frozen during encoding. However, this was the case of the Power Compressor when processing large files, while the LDF Plug-In was rather stable, but it had to manage very large files by default that caused also some problems.
We were happy to find interesting test results of bi-level compression schemes on the web page of the Image Power, Inc.; luckily also with nine tested images2. The images were available in GIF format. We downloaded them and we converted them into bi-level uncompressed TIFF images that were designed to be considered as our source test files.
The Image Power results contained - as a point of reference - also measurements for CCITT G4 and the company's JBIG2 schemes. We repeated the tests for these two schemes on the supposedly identical images of which - after their downloading and conversion - we were not sure enough how much they were identical to those on which the Image Power, Inc., performed the tests.
As to the tests of the CCITT G4 compression scheme, we obtained quite comparable results, while the slight differences might be caused by the fact that our tests ran on slightly - but not critically - modified images. If looking at the graph both CCITT G4 (NL stays for the National Library and IP for Image Power Inc.) curves have similar developments from document to document.
This was not, however the case of JBIG2 compression performed in the Image Power Compressor. Here some results were similar, while the other ones looked differently. This fact might be caused by different pre-processing techniques used by the company in their tests opposite to the technique implemented in their Power Compressor for JBIG2. In our further analyses on other more complicated documents, it was evident that the JB2 format built in the Power Compressor gave unsatisfactory results that will be discussed later.
Nevertheless the CCITT G4 results - and even the JBIG2 results - confirmed
the fact that the downloaded samples are appropriate for our tests. In
only radical difference in JBIG2 compression is observable at the results
concerning the Sample no. 9.
If analysing this sample, we can see that it contains also extremely small (fine) characters in sets of 10, 8, 6, 4, 3, and 2 points. It may be that the quality of pre-processing routines of the encoder used by the company was better than the quality of pre-processing routines used by the encoder built in the available software. As we discovered later, the JBIG2 compression built in the Power Compressor was lossy and gave poor results on documents with small characters.
The differences in results seemed to be explained also by the fact that in case of their another series of tests  the company stated that they had used other tools for simulating JBIG coding scheme.
Furthermore, when we printed out the sample page, we could not discern the smaller sets of characters. It is highly probable that the available encoder was not able to work with them separately; therefore, it encoded bitmaps of these groups of pixels as a whole and the file remained relatively large. This fact seems to be proved also by the fact that at this sample the compression scheme built in the PNG format - good for images - was much more successful than the lossless CCITT G4 and G3 schemes and the result of JBIG2 obtained through the Power Compressor.
This result was very close to the result of classical lossless schemes: 132 KB for JBIG2 compared with 137 KB for CCITT G4 and 145 KB for CCITT G3, while the PNG output was only 77 KB. Also very similar results for both CCITT G3 and G4 confirm this idea, because normally G4 is far better than G3. Maybe this was also the cause of the fact that in another series of tests , the company excluded this sample from tested materials.
The obtained results warned us to be careful when comparing our data with the published results of the company: if doing this, small differences in compared data should be disregarded, while more substantial ones should be analysed. However, our aim was not so much to compare our measurements with other measurements even on the same data; we wanted to have standard monochrome images with different graphic objects (characters, graphs, halftones) to perform our own tests on available bi-level compression schemes.
In conclusion, we found the downloaded samples good for performing our own tests. The appropriateness of our choice of was also strengthened by the fact that it was a set of documents chosen originally by CCITT as reference test documents.
The company's tests contained also values obtained through implementation of special lossy pre-procession techniques3 called Combined Symbol Matching (CSM) and Modified Combined Symbol Matching (MCSM) that gave very interesting results. This fact promised to be interesting because of availability of lossy options also in the DJVU bi-level compression scheme.
The results are shown on the enclosed graph below; they contain also partial
results discussed separately.
It is also interesting to see that G3 was outperformed by PNG in all the tested cases, while G4 only in case of the sample no. 10, which contains halftones and very small characters.
|JB2 is a component part of the DjVu
format. The available software (see above) accepts bi-tonal source files
and it activates in this case only JB2. It is possible to the resolution
of the resulting image and the character of compression. Both of these
parameters can influence the size of the compressed file.
We accepted 300 dpi resolution option, which is the default value. We compressed the source files separately under the lossless, lossy, and aggressive options. The lossless compression outperformed all the other previously discussed schemes.
In comparison with TIFF CCITT G4, it was far better especially in case of texts (samples nos. 4 and 7) and small text and halftones (last sample, here no. 9).
|In comparison with the LDF bi-tonal encoder, the JB2 lossless option was far better also in case of texts (samples nos. 4 and 7). As to the last sample, the difference was not so relevant. This facts seem to say that JB2 disposes of the best textual image pre-processing and processing routines that consists namely in efficient handling of repeated characters. It was successful to do it both in case of Latin and Chinese characters; therefore, it can be said more generally that the quality consists in handling individual pixel blocks. In fact, also the performance of the LDF bi-tonal encoder contained some improvement of this routine face to CCITT G3 and G4 schemes. The results obtained in case of the last sample (here no. 9) prove that the LDF and JB2 encoders have brought significant progresses in handling especially bi-level halftones. The quality of the lossless compression of the DJVU JB2 encoder (a variation to JBIG2) is underlined especially by good and freely available authoring software as well as various other commercial tools. DjVu is also able to handle multipage documents. This quality will be implemented in the version no. 3 of the DjVu™ Shop. The browser plug-in enables Netscape and Internet Explorer users to access and manipulate the djvu files on Internet.|
|Further investigation concerned lossy encoding schemes of the DJVU JB2. As it is evident from the nearby graph, the lossy and aggressive compression options outperformed the lossless encoding within the same format namely in case of textual samples nos. 4 and 7. This signalizes an improved handling of individual pixel blocks represented by characters, indifferently of the fact, whether they are Latin or Chinese. As the density of text diminishes, the prevalence of both lossy schemes over the lossless one weakens or almost disappears, while in case of halftones the results are identical (no. 9). It is also interesting to observe that the differences between the lossy and aggressive options are almost none or negligible. It seems that more complicated source files (documents) will be needed to discover the character of losses in individual options as well as the conditions under which the significance of the aggressive option may become very important face to the lossy one. More above, we stated significant differences between the size of files encoded by these options in case of an old and large page from a 19th century journal.|
In this method individual pixel blocks (in texts in fact images of individual characters) are analyzed in order to find out, which of them and where occur more than once. The principle consists in the fact that each pixel block of the same category is repeated many times on the same page as, for example, the pixel block representing the character 'a' can have a lot of occurrences on a page. If it is so, the bitmap of such a character is encoded only once and it is added to a dictionary of representative pixel blocks. If the encoder finds another pixel block matching the pixel block, which is in the dictionary, this new pixel block is not encoded again. Only its reference to the pixel block from the dictionary is marked as well as the position within the page. Thus, the bitmap of a character is encoded only once, while all its occurrences are given as references to the representative block from the dictionary.
As on scanned textual pages there can be much noise and not all the characters
are imaged by the same number and arrangement of pixels, the encoder must
weigh, whether the match would be acceptable or not. If it is acceptable,
the pixel block is not encoded, it is only substituted by the pixel block
from the dictionary. If the match is not acceptable, the pixel block (character)
is encoded as a bitmap and added to the dictionary. Such a method is called
matching and substitution.
At the neighbouring figure, the upper inscription is from an uncompressed image, while the inscription below has been taken from an image compressed by the Image Power Compressor into their JBIG2 (JB2) format. If we consider the character 'k', we can see that the upper source image contains two slightly different pixel blocks staying for this character. The applied encoder found the second occurrence of this character as matching the first one; therefore, it substituted it with the same pixel block so that the two characters got the same appearance - as seen on the same image below the source text.
However, in case of other characters, there were many misinterpretations in the compressed sample. Some of them did not affect the basic understanding - for example, the last character 'a' - while the other ones changed it as it was, for example, with the second character - that was 'e' - which was matched to something that is rather understood as 'a'. Furthermore, the character 'i' was changed into 'l'.
The source file was of poor quality in this case. It was taken from a scanned
instruction enclosed with a medicament. It was printed with very tiny characters
on normal paper, while the resolution was 300 dpi that was, however, quite
It was interesting in this case to find out, whether also other lossy schemes did the same. As we can see on the following sample, the behaviour of the DJVU JB2 aggressive scheme was better balanced and it did not affect the correct understanding of the text. It is evident that the encoder used a better method, probably that based on the similar principles as those called soft pattern matching in which the pixel block from the dictionary is used as a template to ensure more accuracy during interpretation of characters (pixel blocks). It is a very robust method to small errors5. This file built on this method was slightly larger than then previous file built only on the basis of the pattern matching and substitution method.
However, its readability is quite good, while that of the previous file is seriously affected.
It is evident that the methods called Combined Symbol Matching (CSM) and Modified Combined Symbol Matching (MCSM) mentioned above  are based on similar principles as well as methods used for compressing halftones.
On the results obtained through lossy bi-level compression, some losses can be observed that are obtained especially by flipping pixels. This flipping of pixels is due namely to corrections resulted from the comparison of pixel blocks with pixel blocks from the dictionary. It can prepare good basis for substitution of characters or pixel blocks if the match is considered sufficiently good.
On the comparison of the drawing of a model engine carburettor as obtained
from lossy, lossless, and aggressive compression in DJVU JB2, we can see
several differences caused by flipping pixels.
It is interesting to see also that there is no difference between the aggressive and lossy methods, while there are losses made by these two methods face to the lossless image.
If looking closer, we can distinguish clearly many changes caused by the
fact that individual pixels flipped: from black to white or from white
to black. This is done frequently in order to smooth edges of individual
In this case especially protruding singular black pixels are removed by flipping to white or white pixels flip to black to fill in the edges. Both of these situations can be seen at the nearby picture.
This smoothing procedure is even much more important in case of textual documents, where it can contribute to better matching of encountered characters with the representative character encoded as a bitmap in the dictionary. Smoothing standardizes local edge shapes and it can increase the compression ratio with about 10%.
In case of our tests with compression of scanned pages from the 19th century journals and newspapers, this method proved itself to be very efficient, because the scanned documents offered a larger variety of pixel blocks than common modern printed texts. It was especially on such documents, where differences between lossless, lossy, and aggressive encoding schemes of DJVU JB2 grew substantially. When comparing a sequence of words [Fig. A] from such an old printed text, we could find plenty of cases of smoothed edges. This can be seen, among others, on the character 'o' taken from a larger text of a 19th century journal.
It seems that thanks to this smoothing method very good results are obtained by the lossy DJVU JB2 option face to the lossless encoding: there were 123 KB obtained through lossy encoding compared with 154 KB obtained through lossless encoding.
When we compared the pixel blocks (characters) obtained through lossy and aggressive options, the differences between the shapes of character blocks were none or very few. The same page compressed by the aggressive method, however, had only 109 KB.
To find out, how the aggressive method works in comparison with the lossy one, we took even a larger and more complicated page from a 19th century newspaper. Again we observed very few smoothing differences between lossy and aggressive methods. It seemed that almost everything had been already done by the lossy option. However, it was discovered that the aggressive option had a very strong de-noising feature that could be seen on the sequence of words taken from a processed page of this newspaper [Fig. B]. The aggressive method removed noise around text and between lines and probably thanks to this fact, more compression efficiency was achieved. Also some pixel flipping was observed.
In the graph below, the curves represent how the DJVU JB2 compression becomes more and more efficient face to the lossless option marked by no. 1.
is quite characteristic that the aggressiveness of compression is most
important in case of the old newspaper. Here, the savings are considerable
and they are very significant for slow Internet connectivity. The nearby
image shows differences in absolute savings in kilobytes, while the intersection
of the three curves is the lossy compression result. It is evident from
here that the aggressive method is relevant for transfer of large files.
It is even more important if we realize that the quality of the data issued
from aggressive compression is very good. The only problems could be only
with low resolution files and very small characters.
|However, perhaps more interesting would be to see the relative efficiency of the three DJVU JB2 options on these files. This is shown on the nearby graph relatively to the lossless option. It is evident that the efficiency of the lossy compression is higher in case of modern text, but there is no progress from lossy option to the aggressive one. This is caused by high quality and clearness of the modern text. However, for old documents, the relative savings are not so high: in both cases only ca. 30%, while the aggressive option saves another ca. 10% face to the lossy one. This is probably caused by better smoothing of pixel blocks and their better matching to those from the dictionary and especially by noise removal. Certain role may be also played by quantization of offsets, but here the gain will be very low .|
|We have also observed earlier that DJVU JB2 lossless option is evidently better than the CCITT G4 lossless scheme. Another graph shows this comparison enlarged with comparison with the aggressive JB2 scheme. It can be seen that face to CCITT G4 the JB2 aggressive scheme saves 77% in case of modern text, 43% in case of the old journal, and 45% in case of the old newspaper page. The files received from this aggressive compression are very well readable, while for the normal user and his eyes apparently no inconvenient loss has occurred. Furthermore, it can be said that more critically can be felt losses happened during conversion from reality into only bi-level scheme than those - almost imperceptible - which have taken place during the aggressive lossy compression in DJVU JB2.|
We can see that in both cases JB2 was successful to compress down the file in a very substantial manner. In case of the scanned file there was another additional compression made by the lossy encoder. This was probably done thanks to above-mentioned pre-processing techniques.
As to the computer generated file, the JB2 lossless compression was very radical and it was not improved either by the lossy or aggressive encoders. The explanation is very simple: the file was clean of noise and irregularities; therefore, all the pre-processing and processing techniques could work at their best. The resulting size was very small: only 7 KB.
they post-process the image in certain manner so that the user has the
impression that the image has been improved. This is quite pleasant in
comparison with classical black-and-white viewing. If we look closer to
post-processed characters, we can see that their edges have been filled
in with grey pixels. If such an image is scaled down, we have a very soft
impression of good readability. It disappears usually when we enhance the
text too much, because the illusion works better on smaller characters.
This quality is implemented also in some other plug-ins, for example, Watermark
Web Series that enables to work with TIFF format in Netscape and Internet
Explorer browsers. There is also the ImgGear plug-in developed by the Czech
company Albertina icome Praha that, among other formats, makes possible
work with TIFF CCITT G4 scheme in Netscape. Other plug-ins can be found
on Internet, but they are not so many, while only very few of them are
offered as freeware. It seems that Internet access to bi-level documents
will play an important role in the future; therefore, it makes sense to
work on better compression schemes and better access tools to black-and-white
Also our digital library of old newspapers and journals - from which we took two samples - relies on this scheme and TIFF format. Nevertheless, there are two ready-to-use solutions for document delivery - LDF and DJVU - with better coders of bi-level images that the CCITT G4 is. Especially DJVU for the bi-level image is very promising, because the format is supported by good authoring and viewing tools from which several ones are offered as freeware. This is not the case of LDF or other solutions about which we can only read (e.g. Lightning Strike bi-level encoder by Infinop, Inc.).
It seems that DJVU, as a mixed format, will enjoy of further development: a DjVu encoder for compressing several pages together is about to be launched. Thus, the dictionaries for soft pattern matching will be shared between pages and the compression ratio optimized even more. DjVu is also the format, which has already implemented the emerging JBIG2 standard and which has got at its disposal good tools for work.
It also becomes evident that the future is open for lossy compression schemes for the bi-tonal image. With our tests, we have tried to demonstrate possible results and problems. We have relied on available solutions and software, because we think that the free availability of authoring and viewing tools can contribute to standardization of optimized data formats.